LoopWAM preprint reports robot learning gains from mixed-quality deployment experience

Researchers report that LoopWAM raised robot task success from 23.67% to 43.33% on RoboChallenge Table30-V2 through three rounds of deployment and training. The preprint describes a model that learns from successful actions, suboptimal behavior and failures recorded by itself and other deployed robot policies. LoopWAM retains video evidence across those quality levels while weighting which actions it learns to imitate. Its predictions of physical consequences inform action generation. On a challenging shirt-folding task, two rounds of deployment-driven improvement after training for that task raised success from 13.20% to 55.56%.